Image Inpainting Passive Forensics Based On The Deep Neural Network

At present,more and more digital images are used in social life.The authenticity and integrity of images has received much attention because of the increasing power of image tampering methods.Passive techniques for image forensics is an effective method to detect image tampering.As a popular tampering technique,image inpainting can achieve forgery without any noticeable traces,which poses huge challenges to passive forensics.Passive techniques for image inpainting forensics judges the authenticity and integrity of images by detecting the features of images.Considering that the depth neural network has achieved remarkable achievements in automatically finding the features needed for classification.In this paper,we uses deep neural network to detect image inpainting,and the innovations and contributions are as follows :The proposed network integrates Encoding and Decoding.The Encoding includes convolution layers and pooling layers.The decoding includes deconvolution layers.Firstly,convolution layers extract the features needed for image classification automatically.Pooling layers reduce the number of parameters and improve network robustness.In order to adapt to different size of the input images,the network does not use the fully connection layers.Secondly,the depth neural network tends to learn the features that express the main content of the images and ignore other features.However,the traces left by image inpainting are very weak.The proposed network uses deconvolution to achieve pixel level prediction by recovering feature maps to the input size.The performance of the proposed method is demonstrated on several inpainted images which have tampered areas of different sizes and shapes.TPR up to 96.54%,and FPR can be as low as 0.5%.The average detection time is 2s,which is faster than other existing Passive techniques.The proposed method has good robustness and generalization.The method can resist the interference of JPEG compression operation and zoom operation.